Meta's Chief AI Scientist Yann LeCun, a leading figure in the field of artificial intelligence, is advocating for a shift away from generative AI in pursuit of artificial general intelligence (AGI). AGI represents a level of AI that matches human cognitive abilities, a goal that LeCun believes necessitates a deeper understanding of the world by AI systems rather than mere text generation.
At the recent World AI Cannes Festival, he candidly criticized the current state of machine learning, asserting that “machine learning sucks” in its current form. He emphasized the need for AI systems that can remember, reason, and plan in order to create smarter, more capable technologies. “The future of AI, I tell you, is non-generative. It works for text but doesn’t translate well to broader applications,” LeCun stated.
LeCun is a proponent of joint-embedding architectures (JEPA), which he believes will pave the way for more intelligent AI. Last summer, Meta introduced its first JEPA-based system, I-JEPA, designed to predict missing information rather than just focus on text. “Predicting text is straightforward,” LeCun explained. “However, applying this to the real world presents a much more complex challenge due to the multitude of details involved.”
The ambition to enhance JEPA continues, with plans to extend its capabilities to include video and images. LeCun mentioned that while I-JEPA has not yet utilized a vast dataset, its potential could soon surpass that of DINOv2, Meta’s current computer vision model that employs self-supervised learning.
During the festival, LeCun also shifted the conversation from generative AI to discussing AGI, underlining a critical perspective: “No AI system, no intelligent system, is truly general, including humans. In reality, we struggle in many areas.” He proposed the term “Advanced Machine Intelligence” (AMI) instead of AGI, arguing that while this rebranding is helpful, the journey to achieving such intelligence remains lengthy and complex. He affirmed that while machines will eventually surpass human capabilities in various domains, significant advancements will not occur within the next decade.
The challenges inherent in current AI technologies were underscored by LeCun’s pointed criticism of autonomous vehicles. He remarked, “Any 17-year-old can learn to drive with just 20 hours of practice. Yet, we still lack Level 5 autonomous vehicles without relying on extra tools like sensors and maps.” Level 5 driving signifies complete autonomy with zero human involvement.
LeCun elaborated on the difficulties of defining human intelligence, describing it as specialized, non-linear, and shaped by a rich tapestry of skills and life experiences. He pointed out that relying solely on large language models (LLMs) would obstruct the path toward human-level intelligence. “Much of our understanding comes from our interactions with the world, not just from language,” he explained. “Anyone claiming that AGI is just around the corner is mistaken.”
Moreover, he highlighted that even young children possess a broader experiential base than LLMs, which rely exclusively on textual data. “A child has absorbed 50 times more information than LLMs trained on the entirety of accessible text,” LeCun noted, emphasizing the limitations of current AI training methods.
Though he acknowledged the utility of auto-regressive models, which predict outcomes based on historical data, he criticized their narrow scope, asserting that they represent only a fraction of human knowledge: “They serve useful purposes, but they are not the route to achieving human-level intelligence.”
Looking ahead, LeCun envisions a future where AI assistants mediate every digital interaction for humans. He sees the potential for these assistants to be embedded in devices like smart glasses, a direction that aligns with Meta's innovation strategies. However, he stressed the urgency for AI systems to encompass memory, reasoning, and planning abilities—qualifications that current LLMs lack.
According to LeCun, these future AI assistants must be as diverse as their users, accommodating a variety of languages, cultures, and interests. Achieving this vision will not be feasible for a select few technology companies. “This cannot be accomplished by a small set of firms on the West Coast or in China,” he asserted, suggesting a more collective approach is necessary.
In summary, LeCun believes future AI systems should serve as a vast repository of human knowledge, much like the internet or Wikipedia. “AI has the potential to enhance human intelligence rather than threaten it,” he defended, dismissing alarmist narratives about AI posing an existential risk.
Ranking
